Introduction: Cardiovascular (CV) mortality remains high among adults with diabetes. The impact of behavioral and social factors contributing to this heightened risk is not well understood, and the nonlinear patterns of these determinants can be challenging to capture using traditional linear models. Hypothesis: We hypothesized that explainable machine learning framework based on a tree-based model with Shapley Additive Explanations (SHAP) can capture nonlinear associations between behavioral and social determinants of CV mortality among individuals with diabetes. Methods: Adults with diabetes were identified from the National Health and Nutrition Examination Survey 2007–2018 and linked to the National Death Index. CV mortality was modeled with a tree-based gradient boosting classifier. Feature contributions were assessed using SHAP. SHAP-derived odds ratios (ORs) were obtained by exponentiating differences in mean SHAP values between categories or per-unit increases in continuous variables. Confidence intervals (CIs) were estimated from 1,000 bootstrap samples. Nonlinear associations in continuous variables were modeled using piecewise regression of SHAP values. Results: Among 5,734 adults with diabetes, 516 CV deaths were identified. Of the top 20 contributors in the model, 8 were behavioral or social determinants, accounting for 38% of the model’s explainability. Being U.S.-born was associated with higher odds of CV mortality (OR 1.33, 95% CI 1.33–1.34), as were short (9 h) sleep durations compared with 7–9 h (OR 1.26 and 1.19, respectively). In contrast, recent weight-loss attempts were associated with lower odds (OR 0.68, 95% CI 0.68–0.69). Compared with non-Hispanic Whites, odds were lower among Hispanics (OR 0.81, 95% CI 0.81–0.81). Higher education (college graduate vs less than high school) was associated with reduced odds (OR 0.71, 95% CI 0.71–0.72), as was higher income measured by the income-to-poverty ratio (PIR ≥5 vs <1; OR 0.77, 95% CI 0.77–0.78). Physical activity showed a nonlinear relationship, lowering CV mortality by 0.9% per 10 minutes up to approximately 215 minutes per week, after which further activity yielded minimal benefit. Conclusions: Behavioral and social determinants accounted for a substantial portion of CV mortality risk in adults with diabetes. SHAP-driven interpretability revealed nonlinear relationships, emphasizing modifiable behaviors and social context as key contributors to mortality heterogeneity.
Ahn et al. (Tue,) studied this question.